Tools · Cheminformatics & Structure

Bioactivity Dataset Curator & Splitter

Turn assay exports into clean, reproducible SAR and QSAR datasets with leakage-resistant train, validation, and test splits.

Updated October 1, 2026

On this page

Prices, workflows, and method papers Open in the app

The Bioactivity Dataset Curator turns a raw activity table (a ChEMBL export, an assay spreadsheet) into a dataset you can model on: units normalized, replicates merged, censored values handled, every compound given a stable identity, and splits assigned so the same structure never lands in two of them. Every row it drops is listed with a reason.

How it works

Each row goes through the same steps:

Value: a number, optionally with a relation (<10, >=5, or a relation column).
Endpoint: from the endpoint column (or default_endpoint), normalized (ic50 → IC50). Set endpoints to keep only some.
Units: pM to M concentrations become p_activity = −log₁₀(molar); log-scale values pass through. On the p-scale relations flip: IC50 < 10 nM becomes p > 8. Values outside 0 to 14 are rejected as implausible.
Censoring (censored_policy): drop bounds, treat them as_threshold, or keep them apart from exact values.
Identity: RDKit parses the SMILES and the standard InChIKey becomes compound_id. Salts and tautomers are not standardized; run Molecule Standardizer first if you need that.

Then rows for the same compound and endpoint (and assay, with group_by_assay) are aggregated by median or mean. If replicates disagree by more than max_replicate_spread log units (default 1.0), the whole group is rejected as a conflict rather than averaged. Every input row ends up either in a curated record or in the rejections file.

Splits

Splits are assigned per compound, so a structure never appears in two of them:

scaffold (MoleculeNet-style, deterministic): Bemis–Murcko scaffold groups, largest first, the largest always in train. The usual choice for estimating how a model will do on new chemotypes.
random: a seeded shuffle.
time: by each compound's earliest measurement date (needs a date or year column).

split_fractions is [train, valid, test] and must sum to 1.

Inputs

One activity table (CSV, TSV, JSON, a list of rows, or an artifact), up to 50,000 rows. Name at least the SMILES and value columns; units, relation, endpoint, assay, compound id, and date columns are picked up from common header names (ChEMBL's are recognized first). A curated table from an earlier run can be curated again without remapping.

Outputs

File	Contents
`curated.csv`	One record per compound and endpoint: `p_activity`, value in nM, number of measurements and their spread, censoring, scaffold, and split
`splits.csv`	Each compound's scaffold, split, and date
`rejections.csv`	Every dropped row with its reason
`molecules.smi`	Each compound's SMILES and InChIKey
`summary.json`	Counts, rejection reasons, resolved columns, and warnings

Feed the curated table to Matched Molecular Pairs or Free Wilson Analysis, or train a model on it with MPNN Models.

Run it from the API

Submit with Submit a job and the job_type below. Price it first with Estimate job reservation cost: submitting reserves that amount from your wallet, and the charge settles at the actual runtime.